Has Google's AI watermarking system been reverse-engineered?

TL;DR

A developer claims to have reverse-engineered Google DeepMind's SynthID watermarking system, showing AI-generated watermarks can be removed using signal processing and 200 Gemini images without proprietary access. Google disputes the claim's validity, but the open-sourced code invites verification and raises serious questions about watermarking as a reliable AI governance tool.

• •Developer claims successful reverse-engineering of SynthID watermarking system
• •Method uses basic signal processing + Gemini images; requires no proprietary access
• •Open-sourced code invites verification and scrutiny of watermarking robustness

Generated with AI, which can make mistakes.